The Workforce Selection Effects of Training with AI: Evidence from a Field Experiment
Emaad Manzoor, Hongye Sun, Kun Peng, Vrinda Kadiyali · Cornell University, University of Utah, Nanjing University · 11 Nov 2026
Source finding
Two weeks after training, retention was 13.1 percentage points higher after AI practice than after peer practice. The comparison covers complete training formats, not the language model alone.
Our interpretation
Training formats can change who stays. Assess retention, quality and business outcomes separately.
Method and evidence quality
Random assignment within twelve recruitment cohorts; shared classroom teaching and allotted practice time. Administrative retention records. Strong local short-term comparison, not a general market rate.
Limits and interests
One Chinese firm, short follow-up and a working paper not peer reviewed. AI practice was private without explicit feedback; peer practice included supervision and feedback. The comparison does not isolate the language model. More calls do not prove higher revenue. No separate SME effect.
What businesses can do
Assess new training formats beyond practice performance. Compare who stays, work quality and actual sales outcomes.
Best reading entry
Table 5 and sections 2.2 and 3.4; Table 1 describes groups.
Original
The Workforce Selection Effects of Training with AI: Evidence from a Field Experiment
Try it in your business
Our application idea, separate from the study finding
Our application idea, not an observed study case: An SME sales team trials AI role-play for new recruits. Alongside conversation quality, it checks whether and which recruits remain after onboarding.
First steps, measures and team check
Who this is for
Sales lead · Learning lead · Business owner · Sales · Onboarding
Relevant to recurring sales onboarding. A manufacturer with about 600 salespeople is not a representative SME sample; test local effects.
Misuse to avoid
Misuse to avoid: Treating more practice calls as better sales performance or assigning training solely on personality-test results.
Steps for this week
- Define conversation quality and retention after two and six weeks as separate measures
- Trial two formats with the same learning goals and document practice time, supervision and feedback differences
- Assess retention, conversation quality and actual sales separately; adapt training with participants
What to measure
Share of original recruits still employed at the agreed follow-up date
Assess conversation quality independently with shared criteria; record completed sales and revenue separately
Keep the original group in the denominator after departures. Document group and supervision differences. Do not borrow a target rate from this experiment
Team check
- Who remains employed after two and six weeks?
- Do supervision or feedback differ between training formats?
- Do conversation quality and completed sales also improve?
Evidence and publisher interests
One Chinese firm, short follow-up and a working paper not peer reviewed. AI practice was private without explicit feedback; peer practice included supervision and feedback. The comparison does not isolate the language model. More calls do not prove higher revenue. No separate SME effect.
Academic authors collaborating with an anonymous medical equipment firm. The AI practice environment was developed with sales managers. No direct AI product sale disclosed in the reviewed paper; this is not a claim of complete absence of conflicts.